DAOS-16876 vos: skip DTX record when load partial committed DTX

It is used to handle existing partial committed DTX records that were
generated when ran as DAOS-2.6.3-rc{1,2} to avoid repeated DTX commit
after engine restarted.

To be safe, it is required for the user/admin to explicitly set server
side environment variable "DAOS_SKIP_OLD_PARTIAL_DTX" when restart the
engine that has ever run as DAOS-2.6.3-rc{1,2}. If the user has never
used such version, such as directly upgrading from 2.6.1 to 2.6.3-rc3
or newer, then please do NOT set this environment variable.

Signed-off-by: Fan Yong <[email protected]>

src/vos/vos_common.c

@@ -716,6 +716,14 @@ vos_mod_init(void)
 	d_getenv_bool("DAOS_DKEY_PUNCH_PROPAGATE", &vos_dkey_punch_propagate);
 	D_INFO("DKEY punch propagation is %s\n", vos_dkey_punch_propagate ? "enabled" : "disabled");
 
+	/*
+	 * NOTE: It is used to skip old partial committed DTX records that were generated when
+	 *	 ran as DAOS-2.6.3-rc{1,2}. If the user has never used such version, please do
+	 *	 NOT set this environment variable.
+	 */
+	d_getenv_bool("DAOS_SKIP_OLD_PARTIAL_DTX", &vos_skip_old_partial_dtx);
+	D_INFO("%s old partial committed DTX record\n", vos_skip_old_partial_dtx ? "Skip" : "Keep");
+
 	vos_agg_gap = VOS_AGG_GAP_DEF;
 	d_getenv_uint("DAOS_VOS_AGG_GAP", &vos_agg_gap);
 	if (vos_agg_gap < VOS_AGG_GAP_MIN || vos_agg_gap > VOS_AGG_GAP_MAX) {

src/vos/vos_dtx.c

@@ -41,6 +41,8 @@
 			    DAE_EPOCH(dae));						\
 	} while (0)
 
+bool vos_skip_old_partial_dtx;
+
 static inline uint32_t
 dtx_umoff_flag2type(umem_off_t umoff)
 {
@@ -706,7 +708,7 @@ dtx_rec_release(struct vos_container *cont, struct vos_dtx_act_ent *dae, bool ab
 		 * If it is required to keep the active DTX entry, then it must be for partial
 		 * commit. Let's mark it as DTE_PARTIAL_COMMITTED.
 		 */
-		if ((DAE_FLAGS(dae) & DTE_PARTIAL_COMMITTED))
+		if (DAE_FLAGS(dae) & DTE_PARTIAL_COMMITTED)
 			return 0;
 
 		rc = umem_tx_add_ptr(umm, &dae_df->dae_flags, sizeof(dae_df->dae_flags));
@@ -3235,6 +3237,9 @@ vos_dtx_act_reindex(struct vos_container *cont)
 			dae->dae_need_release = 1;
 			D_INIT_LIST_HEAD(&dae->dae_link);
 
+			if (vos_skip_old_partial_dtx && DAE_FLAGS(dae) & DTE_PARTIAL_COMMITTED)
+				DAE_REC_CNT(dae) = 0;
+
 			if (DAE_REC_CNT(dae) > DTX_INLINE_REC_CNT) {
 				size_t	size;
 				int	count;

src/vos/vos_internal.h

@@ -149,6 +149,7 @@ extern uint32_t vos_agg_gap;
 
 extern unsigned int vos_agg_nvme_thresh;
 extern bool vos_dkey_punch_propagate;
+extern bool vos_skip_old_partial_dtx;
 
 static inline uint32_t vos_byte2blkcnt(uint64_t bytes)
 {